Not Hackers, Just an Inside Mistake
It was 26 degrees in Shanghai today, with a clear blue sky. Reading this news in such weather, I was stunned for a good while.
Anthropic revealed the trump card of its most powerful AI themselves—not at a press conference, but because their backend configuration was botched.
On March 26, a security researcher casually browsing discovered that Anthropic’s Content Management System (CMS) had nearly 3,000 files completely open to the public internet. No authentication, no passwords, just direct access. Inside were internal blog drafts, PDFs, images, detailed arrangements for a CEO-only event—and the complete introductory materials for their next-generation flagship model, Claude Mythos.
Anthropic quickly locked down permissions after being contacted, but it was too late. The files had already spread across security forums and social media.
To put it bluntly, someone forgot to set the permission to “private” when uploading the files. That’s it. Just one switch.
How Absurd is the Number 10 Trillion?
In the leaked materials, the most discussed detail is the parameter scale: 10 trillion.
I previously read that the human brain has about 100 trillion synaptic connections, and an AI model’s “parameters” roughly correspond to this concept. When GPT-3 was released in 2020 with 175 billion parameters, people already thought it was incredible. GPT-4 in 2023 had roughly 1.8 trillion. Now, Mythos has jumped to 10 trillion.
From GPT-3 to Mythos, this number has multiplied by nearly 57 times.
However, it’s worth noting that 10 trillion parameters is about 1/10th the number of synapses in the human brain. So there’s no need to think AI has “surpassed the human brain” just yet—the human brain can run on a single sandwich, consuming about 20 watts of power. Server farms are a completely different story.
It’s certainly impressive, but it hasn’t quite reached a god-like status.
Interestingly, in the internal documents, Anthropic described Mythos as having “capabilities in cybersecurity that far exceed any other AI model,” and specifically mentioned “unprecedented cybersecurity risks.” This sentence was written by their own people, describing what they built.
By the way, the leaked documents also mention a new model tier called “Capybara.” Anthropic naming their internal product tiers after capybaras—this detail made me love them for a second.
Have Predecessors Made Similar Mistakes?
Actually, this kind of “tech company digging its own grave” story is not a first in the AI circle.
Just a few days after the Mythos leak, another incident happened at Anthropic: someone discovered that the npm package for their development tool, Claude Code, hid a 59.8MB source map file containing a URL pointing to a Cloudflare R2 bucket—and that bucket was completely public, with no authentication. Pulling from that link revealed a full 512,000 lines of TypeScript source code, 1,906 files, and 44 hidden feature flags. Oh, and an embedded virtual pet, Tamagotchi.
Two different types of configuration mistakes, happening at the same company, less than a week apart.
People seriously debated online: was this an accident, or the most brilliant PR stunt in history? I think it was likely an accident—but the effect of this accident was far more viral than any press conference.
Anthropic Has Always Been a Bit Unique
Anyone slightly familiar with Anthropic’s origins knows it is inherently a story of a “breakaway.” Founders Dario Amodei and his sister Daniela Amodei, along with several other core members, were originally from OpenAI. They left in 2021 to found Anthropic, with the core philosophy of building “safer AI.”
And then this safety-focused company built a model described in its own internal documents as bringing “unprecedented cybersecurity risks.”
I have no intention of mocking them—people doing technical research are more sensitive to safety issues, which is why they use such wording internally. But this contrast is indeed quite intriguing.
My biggest takeaway after reading all this: this accidental leak makes the company feel more real than any meticulously planned press conference could. Perfect PR drafts make people wary, but a single unchecked configuration box makes you feel—oh, they’re just ordinary humans too.
Oh, and one more thing: Mythos is currently only open to “select early users,” meaning most people won’t be able to use it for now. The best phrase to understand this situation is probably the three leaked words—
Capybara Tier.
References:
- Anthropic Leaked Its Own CMS: 3,000 Unpublished Claude Files
- Claude Mythos (Opus 5) Leaked: What We Know So Far
- Spud And Mythos: New Models Break On The Shore Of 2026 – Forbes
- The Great Claude Code Leak of 2026 – DEV Community
- Frontier language models have become much smaller – Epoch AI
—— Lyra Celest @ Turbulence τ.
